For network service providers, a key consideration in network design and management is the appropriate allocation of access capacity and network resources between traffic originating from VPN customer sites and traffic originating from outside the VPN (e.g., from the Internet or other VPNs). This consideration is particularly significant with respect to the traffic of VPN customers whose subscription includes a Service Level Agreement (SLA) requiring the network service provider to provide a minimum communication bandwidth or to guarantee a particular Quality of Service (QoS). Such service offerings require the network service provider to implement a network architecture and protocol that achieve a specified QoS and ensure sufficient access capacity and network resources are available for communication with other VPN sites separate from communication with hosts that are not part of the VPN.
In Internet Protocol (IP) networks, a straightforward approach to achieving QoS and implementing admission control comparable to that of connection-oriented network services, such as voice or Asynchronous Transfer Mode (ATM), is to emulate the same hop-by-hop switching paradigm of signaling resource reservations for the flow of IP packets requiring QoS. In fact, the IP signaling standard developed by the Internet Engineering Task Force (IETF) for Integrated Services (Intserv) adopts precisely this approach. As described in IETF RFC 1633, Intserv is a per-flow IP QoS architecture that enables applications to choose among multiple, controlled levels of delivery service for their data packets. To support this capability, Intserv permits an application at a transmitter of a packet flow to use the well-known Resource ReSerVation Protocol (RSVP) defined by IETF RFC 2205 to request a desired QoS class at a specific level of capacity from all network elements along the path to a receiver of the packet flow. After receiving an RSVP PATH message requesting a resource reservation and an RSVP RESV message confirming resource reservation from an upstream node, individual network elements along the path implement mechanisms to control the QoS and capacity delivered to packets within the flow.
FIG. 1 illustrates the implications of utilizing a conventional Intserv implementation to perform admission control. As shown in FIG. 1, an exemplary IP network 10 includes N identical nodes (e.g., service provider boundary routers) 12, each having L links of capacity X coupled to Customer Premises Equipment (CPE) 14 for L distinct customers. In a per-flow, connection-oriented approach, each node 12 ensures that no link along a network path from source to destination is overloaded. Looking at access capacity, a per-flow approach is able to straightforwardly limit the input flows on each of the ingress access links such that the sum of the capacity for all flows does not exceed the capacity X of any egress access link (e.g., Link 1 of node 12a). A similar approach is applicable to links connecting unillustrated core routers within IP network 10.
Although conceptually very simple, the admission control technique illustrated in FIG. 1 has a number of drawbacks. Most importantly, Intserv admission control utilizing RSVP has limited scalability because of the processing-intensive signaling RSVP requires in the service provider's boundary and core routers. In particular, RSVP requires end-to-end signaling to request appropriate resource allocation at each network element between the transmitter and receiver, policy queries by ingress node 12b-12d to determine which flows to admit and police their traffic accordingly, as well as numerous other handshake messages. Consequently, the processing required by Intserv RSVP signaling is comparable to that of telephone or ATM signaling and requires a high performance (i.e., expensive) processor component within each boundary or core IP router to handle the extensive processing required by such signaling. RSVP signaling is soft state, which means the signaling process is frequently refreshed (by default once every 30 seconds) since the forwarding path across the IP network may change and therefore information about the QoS and capacity requested by a flow must be communicated periodically. This so-called soft-state mode of operation creates an additional processing load on a router even greater than that of an ATM switch. Furthermore, if the processor of a boundary router is overloaded by a large number of invalid RSVP requests, the processor may crash, thereby disrupting service for all flows for all customers being handled by the router with the failing processor.
In recognition of the problems associated with implementing admission control utilizing conventional Intserv RSVP signaling, the IETF promulgated the Differentiated Services (Diffserv or DS) protocol defined in RFC 2475. Diffserv is an IP QoS architecture that achieves scalability by conveying an aggregate traffic classification within a DS field (e.g., the IPv4 Type of Service (TOS) byte or IPv6 traffic class byte) of each IP-layer packet header. The first six bits of the DS field encode a Diffserv Code Point (DSCP) that requests a specific class of service or Per Hop Behavior (PHB) for the packet at each node along its path within a Diffserv domain.
In a Diffserv domain, network resources are allocated to aggregates of packet flows in accordance with service provisioning policies, which govern DSCP marking and traffic conditioning upon entry to the Diffserv domain and traffic forwarding within the Diffserv domain. The marking (i.e., classification) and conditioning operations need be implemented only at Diffserv network boundaries. Thus, rather than requiring end-to-end signaling between the transmitter and receiver to establish a flow having a specified QoS, Diffserv enables an ingress boundary router to provide the QoS to aggregated flows simply by examining and/or marking each IP packet's header.
Although the Diffserv standard addresses Intserv scalability limitation by replacing Intserv's processing-intensive signaling with a simple per packet marking operation that can easily be performed in hardware, implementation of the Diffserv protocol presents a different type of problem. In particular, because Diffserv allows host marking of the service class, a Diffserv network customer link can experience a Denial of Service (DoS) attack if a number of hosts send packets to that link with the DS field set to a high priority. It should be noted that a set of hosts can exceed the subscribed capacity of a Diffserv service class directly by setting the DSCP or indirectly by submitting traffic that is classified by some other router or device to a particular DSCP. In Diffserv, an IP network can only protect its resources by policing at the ingress routers to ensure that each customer interface does not exceed the subscribed capacity for each Diffserv service class. However, this does not prevent a DoS attack.
FIG. 2 depicts a DOS attack scenario in an exemplary IP network 10′ that implements the conventional Diffserv protocol. In FIG. 2, a number of ingress nodes (e.g., ingress boundary routers) 12b′-12d′ each admit traffic targeting a single link of an egress node (e.g., egress boundary router) 12a′. Although each ingress nodes 12′ polices incoming packets to ensure that customers do not exceed their subscribed resources at each DSCP, the aggregate of the admitted flows exceeds the capacity X of egress Link 1 of node 12a′, resulting in a denial of service to the customer site served by this link.